OpenAI’s GPT‑5.6 family is presented as a way to lower the cost of agentic applications while preserving frontier-level capabilities. The release pairs smaller models—Luna and Terra—with API primitives intended to keep reasoning coherent across long workstreams and to move deterministic work into code.
Model selection and cost trade-offs
According to published results, Luna and Terra can often match the performance of larger GPT‑5.4 and GPT‑5.5 variants for many production tasks at much lower cost. Hypha reported that “Luna keeps 98% of GPT‑5.5’s extraction accuracy at one‑eighteenth the cost,” and Browser Use said Luna completed 78% of 106 difficult browser tasks for about $14, versus a current SOTA model at 80% for roughly $235.
On the BrowseComp benchmark, GPT‑5.5 (Extra High) scored 84.36% at a cost of $33.27, while GPT‑5.6 Luna (Extra High) scored 84.04% at a cost of $1.33. PlayerZero reported using Luna for code retrieval and decision modeling workloads, claiming a 64% reduction in inference costs, a 90% cut in response time, and a five‑point F1 improvement for a key task.
Responses API features for agent efficiency
The Responses API additions highlighted include retained reasoning and native compaction, multi‑agent orchestration, programmatic tool calling, and longer prompt cache windows. Retained reasoning and compaction allow previously performed work to be persisted and compressed across turns; OpenAI’s tests show this can increase effectiveness while reducing token use. For example, on ARC‑AGI‑3, GPT‑5.6 Sol moved from 13.3% with a standard harness to 38.3% after enabling retained reasoning and compaction, while using roughly six times fewer output tokens.
Programmatic Tool Calling enables the model to emit JavaScript that orchestrates tools and processes outputs outside the model’s context window so the model focuses on judgment. Rogo reported that using this feature matched rubric quality while using 21% fewer input tokens.
Prompt caching now has a minimum TTL of 30 minutes and supports deterministic cache breakpoints within a model’s context window. Ploy said adding cache breakpoints and workspace keys to a 29,000‑token prompt cut uncached input by 28%.
Reports from builders
Startups quoted in the guide attribute cost and latency improvements to a mix of model selection and the new API controls. Izzy Miller, AI Research Lead at Hex, said GPT‑5.6 at low reasoning effort “knew when the data just wasn’t there, didn’t chase bad leads, and got to the right answer with fewer tokens.”
Guide authors are Samarth Madduru, Prashant Mital, Dave Leo and Julien Reiman. The guide was published August 13, 2026 and documents startup experiences with GPT‑5.6 in production testing.
Original source: OpenAI News